No keyboard swap, no separate app to paste out of. Ramblr floats a single button over whatever you're already doing — Slack, Discord, a code editor, an email — and inserts your words directly into the focused field the moment you're done talking.

• Tap, ramble, tap. One button, everywhere. No mode switching, no separate transcript to copy-paste.


• Fully offline if you want it. On-device transcription and on-device cleanup — nothing ever has to touch the network.


• Or route it through the best cloud models. OpenAI and Gemini for transcription; OpenAI, Anthropic, and Gemini for cleanup — chained in whatever fallback order you configure, your own API keys, no relay server in between.


• Fastest of both worlds. Instant on-device transcription, cloud-grade cleanup — dictation feels instant and still reads like you actually meant to write it.


• Cleanup styles that match how you actually talk. Formal, Casual, Notes and lists, Email, Concise — or write your own. Ramblr strips the "ums," collapses your self-corrections, and keeps your meaning, not just your words.


• It learns your vocabulary. Feed it your project names and jargon once; cleanup stops mangling them forever.

Ramblr uses Android's Accessibility Service for one narrow reason: to insert dictated text into the currently focused text field across apps. It does not replace your keyboard. It does not run background automation. It only acts after you explicitly tap the overlay button.

Five ways to start dictating: the floating icon (default), the optional Ramblr Voice keyboard, Android's system accessibility button or a volume-key hold, a Quick Settings tile, or by selecting existing text in any app and tapping "Ramblr" to clean it up. Two of those — the voice keyboard and the select-text integration — work with the Accessibility Service turned off entirely. Ramblr Voice is a voice-only input method, not a full replacement keyboard: it has no letter, number, or symbol rows.

Network connections: Ramblr makes no network connection on its own. Every connection it can make is listed here:

• Speech and cleanup model downloads (on request). On-device transcription needs a speech model, which is downloaded only when you pick one — from github.com (the sherpa-onnx project's release files) or huggingface.co. Optional on-device cleanup and optional silence auto-stop download their models the same way, only when you turn those features on. Nothing is downloaded until you ask for it, and models are verified against a known SHA-256 before being installed.


• Optional cloud transcription and cleanup (off by default). Only if you enable it and supply your own API key. Requests go straight from your phone to the provider you chose (OpenAI, Anthropic, Google Gemini, or a self-hosted OpenAI-compatible endpoint you point it at) — there is no Ramblr server in between.


• Cloud model list (only while setting up a cloud provider). When you open the cloud provider setup screen, Ramblr refreshes its list of available cloud models — which models each provider offers and their prices — from a public file on gist.githubusercontent.com, at most once a day. This exists so a newly released cloud model can be offered without shipping an app update. It is never fetched if you don't use cloud providers, it carries no identifying information, and a bundled list is used if the fetch fails.

There is no analytics, no telemetry, no crash reporting, and no advertising of any kind. This F-Droid build also contains no update checker: it never contacts GitHub to look for new versions, and updates come from F-Droid like any other app.

Privacy: combine local transcription with local cleanup and nothing ever leaves the device. There is no backend for this app — in every cloud mode, requests go straight from your phone to the provider using your own API key, with no relay server in the middle. Cl